DeepRacer service is a platform for developers and enthusiasts alike to get hands-on with reinforcement learning. This is a great starting point for experimenting with different Machine Learning reinforcement models, training them on DeepRacer 3D racing simulator, and deploying these trained models on DeepRacer to get a real world feel.
Many people would think, at first glance, it is a novel toy to play with that after some time will become useless. On the contrary, it is an innovative approach initiated by AWS to fill the gap being left wide due to the lack of Data Scientists. The adoption of Artificial Intelligence technology in a wider commercial domain is suffering, as the cost to hire experts is too high due to the lack of trained human resources. Hence the next best solution is to make the science more accessible by providing tools to the developers. These tools can be used by developers to forecast, improve and effectively integrate with different business models without having an extensive data science background.
Artificial Intelligence, a short introduction
The term Artificial Intelligence (AI) was first coined in 1956 at a Dartmouth College in Hanover, New Hampshire. Artificial Intelligence can be attributed to the branch of computer science focused on enabling digital computers or computer-controlled robots to perform tasks generally associated with intelligent beings. Since its inception in 1956, the field of AI has been progressing over the years. Though it is only in the last 15 years, AI has become popular thanks to the improvements in computers (computing power, GPUs, & storage devices) and its accessibility via cloud-based computing, huge amount of accessible data and the advanced algorithms to train and test these models.
Machine Learning, a subset of AI
Machine Learning (ML) is a subset of much wider field of Artificial Intelligence. Machine Learning at its root is the idea that machines can learn from past data points using analytical models to predict the future behavior of the system for which it is trained. It does this by identifying patterns with minimal human interactions. Within ML, there is a concept of reinforcement learning which in essence is learning via positive reinforcement. It is the same way you would train your pet, reinforcing good behavior with positive feedback when a certain objective is reached.
Schematic depiction of AI, ML and Deep Learning relationships
The serious objective of the Deepracer
In the case of DeepRacer, the objective is to autonomously drive through the track without deviation and to complete this task in the shortest time possible. The exciting part is that the task of building the model and training it for autonomous driving can be orchestrated from a single AWS service.
The AWS DeepRacer stitches together various AWS services to provide a single point control:
- ·The reinforcement model is trained in AWS Sagemaker over a simulated race track. The simulated race track has the same dimensions as the track in the real world.
- This racing simulator is provided by AWS Robomaker.
- Amazon Kinesis Video stream provides real-time footage of virtual simulation and the effects of positive reinforcement constant.
- AWS CloudWatch captures the logs with respect to different models being trained and the different metrics a developer want to see.
- Last but not the least Amazon Simple Storage Service (S3) is where the model is stored, and versioning can be enabled to keep track of various versions of the model.
The DeepRacer can be driven in two configurations:
· Manual driving mode, where the car can be driven via a computer over local Wifi and can be steered through the AWS DeepRacer control page.
· And Autonomous driving mode, where a model is trained for a track and then steers DeepRacer around the track. This anticipated service was launched in May 2019 and is available online.
For the Machine Learning enthusiasts, AWS launched DeepRacer League where the players will race against each other in autonomous driving mode. The team at Xebia is very excited to try the trained models against others during the league and learn as we step into the next domain of cloud-based computing with AWS.
Personally, I will be investing my time in making complex reinforcement models and trying to find ways to train and test these models with DeepRacer. On the other hand, it could be very interesting to train a model for obstruction recognition and finding an alternative route around the said obstruction. The possibilities are endless and the future in Machine Learning is exciting.